subseasonal forecast of surface water conditions in ......1 geophysical fluid dynamics laboratory,...
TRANSCRIPT
1 Geophysical Fluid Dynamics Laboratory, NOAA2 Atmospheric & Oceanic Sciences (AOS), Princeton University3 University of California Santa Cruz and NOAA Southwest Fisheries Science Center4 Institute of Marine Sciences, University of California Santa Cruz5 NOAA National Marine Fisheries Service, Northeast Fisheries Science Center
Subseasonal forecast of surface water conditions in Chesapeake Bay using a
hybrid approach
F.G. Taboada1,2*, K. Dixon1, B. Muhling3, D. Tommasi4,
M.-J. Nath1, V.S. Saba1,5, G. Vecchi2 & C. Stock1
1
S2 - From prediction to projection: the role of
seasonal to decadal forecasts in a changing
climate
*E-mail: [email protected]
Key ecosystem services but subject to important human impacts:
- nutrient cycling, food provision, cultural & recreational uses
[~30K USD ha-1 yr-1, rank 3rd]
- eutrophication, overexploitation and climate change
Challenging environment for resource managers
- multiple drivers and multiple sources of environmental variability
• Streamflow (watershed), atmosphere, coastal ocean (tides)
• Scale markedly differs from target of Global Climate Models and Predictions
Can S2S2D improve management of transitional water bodies in a changing climate?
McLusky & Elliot 2004; Costanza et al 2014 Global Environ Change; Tommasi et al. 2017
Estuarine ecosystems
2
Northeast coast of USA
Huge watershed [~166534 km2]
More than 18 million people living around
- Pollution (eutrophication)
- Climate change (sea level rise, warming)
- Overexploitation (fisheries and aquaculture)
Data for 2016, Chesapeake Bay Program
Chesapeake Bay
3
Northeast coast of USA
Huge watershed [~166534 km2]
More than 18 million people living around
Area 11000 km2
Volume 74.4 km3
- More than 150 tributaries
- Total inflow ~2300 m3 s-1
- Susquehanna river ~50%
Chesapeake Bay Program; USGS Water Services
Chesapeake Bay
4
James
Northeast coast of USA
Huge watershed [~166534 km2]
More than 18 million people living around
Area 11000 km2
Volume 74.4 km3
- 300 km long
- 25 km wide [5-60 km]
- 6.5 m depth [up to 50 m in central channel]
Chesapeake Bay Program; USGS Water Services
Chesapeake Bay
5
You are
here
300 k
m
25 km
Long term changes associated with climate change
Muhling et al 2017 Estuaries and Coasts; Muhling et al 2017 GeoHealth
Background – decadal to century long
6
Projected increase in the
mean concentration of
Vibrio parahaemolyticus in
oyster during summer
Projected changes in temperature and
salinity in surface waters
GCM atmospheric conditions
Statistical Downscaling (ESD)
Regression Trees
temperature salinity
Initial vs boundary conditions
Adapted GCMs configurations
- Data assimilation
- Limitations [e.g. precipitation]
Subseasonal to Seasonal (S2S)
7
Skill for temperature and
precipitation averaged for leads
between 0 and 9 months using
GFDL FLOR forecast system
Jiao et al 2015 J Clim
Thomas Point buoy TPLM2
National Data Buoy Center C-MAN station;
- More than 30 years of daily data
- Air temperature [tas] and sea surface temperature [tos]
CTD data
Sparse but extensive [>40K obs]
- 92 fixed stations, ~ fortnightly/monthly
- Sea surface temperature [tos] and salinity [sos]
Three main contributors;
- Chesapeake Bay Program
- University of Maryland
- Smithsonian Institution
Data – 1986-2015 (30 years)
8
US Coast Guard, P. Milnes
Station data
9
Raw and anomaly time series plots Seasonality Variogram
Raw and anomaly time series plots Seasonality Variogram
Time lag
1. Forecast-oriented Low Ocean Resolution (FLOR)
- NOAA-GFDL Climate Model 2.5
- Data assimilation + HR [50 km atmosphere]
2. Framework for Unified Downscaling of GCMs Empirically (NOAA-GFDL FUDGE)
- Quantile mapping of FLOR forecasts
3. Hierarchical Bayesian spatiotemporal model (HBSTM)
- Initial condition; reconstruct spatial tos and sos anomaly fields
- Forecast anomaly evolution taking advantage of FLOR forecasts and covariates
Vecchi et al. 2015 J Clim; Dixon et al. 2017 Clim Change; Cressie & Wikle 2013
Three component hybrid framework for S2S
12
FLOR anomalies could improve tos forecasts
Downscaled [QMap] to match Thomas Point data and transformed to tos anomalies
Framework for S2S – covariates
13
TPLM2 C-MAN
daily tas
FUDGE
ESD-QMFLOR daily forecast
[1 month lead tas]
1 month lead
ESD-FLOR
daily tas
1 month lead
ESD-FLOR
weekly tos
Surface water temperature
anomaly @ Thomas Point
Past river discharge can inform changes in salinity
Inability to predict precipitation but mechanistically consistent
[residence time ~180 days]
Du & Shen 2016 J Mar Syst
Framework for S2S – covariates
14
River discharge
database
sos CTD time series
@ ‘fixed array’
averaged flow
prev. 3 months
weekly Q
Susquehanna river
discharge @ Conowingo
Candidate river?
Integration period?
Combine different data sources;
- Covariates provide info at a single point
- Target observations irregularly distributed
Hierarchical spatiotemporal model;
- Based on Wikle et al. 1999 Environ Stats
- Reconstruct underlying field
- Project short term evolution
Cressie & Wikle 2013
Framework for S2S – hierarchical model
16
1. Weekly evolution of tos or sos in a 10 km grid
2. Space-time decomposition based on spatially varying;
a. Mean [𝜇 𝑠 ]
b. Seasonal cycle [𝑆 𝑠, 𝑡 ]
c. Covariate effect [𝜈 𝑠 𝑁 𝑡 ]: none, FLOR or Susquehanna
d. Local decay and (fixed) neighbor effects (VAR) [𝑋 𝑠, 𝑡 ]
3. Spatial smoothness constrain (Gaussian random fields)
4. Assessment based on sequential fitting (2011-2015)
and one-step ahead forecasts (1 month lead)
Cressie & Wikle 2013; Wikle et al. 1999 Environ Stats
Framework for S2S – hierarchical model
17
𝑍 𝑠, 𝑡 ~ 𝑁𝑜𝑟𝑚𝑎𝑙 𝑀 𝑌 𝑠, 𝑡 , 𝜎𝜖
𝑌 𝑠, 𝑡 = 𝜇 𝑠 + 𝑆 𝑠, 𝑡 + 𝜈 𝑠 𝑁 𝑡 + 𝑋 𝑠, 𝑡
𝑋 𝑠, 𝑡 = 𝐻 𝑋 𝑠𝑁𝑁, 𝑡 − 1 + 𝛾𝑡−1,
𝛾𝑡 ~ 𝑁𝑜𝑟𝑚𝑎𝑙(0, 𝜎𝛾)
Results – covariate effects
18
River discharge decreases
salinity in lower bay but nil
in upper bay [always fresh]
Temperature anomalies
are more similar to ESD-
FLOR close to TPLM2
Results – persistence
19
Reinforces the notion that
river anomalies are more
persistent
(note that the model includes
the covariate effect)
Results – example predictions
20
Works relative well in many
cases …
[… others not!]
Results – monthly correlation skill
21
Correlation between predictions and observations,
different months for 2015
NOTE the change of scale between sos and tos panels
Hybrid approach improves predictability wrt either approach independently
(statistical or dynamic); take the best of the two worlds
Downscaling highlighted again as key approach to solve the change of support
problem and to account for model biases
Potential applications towards decreasing the impacts associated with
- marine pathogens (Vibrio) and jellyfish blooms
- eutrophication and anoxia
Explore potential extension to other transitional water bodies
Summary and prospects
22
23
With special thanks to
Carlos Gaitán [Arable Labs]
Andrew Ross [NOAA GFDL/AOS]
Xiaosong Yang [NOAA GFDL/UCAR]
Data sources
CTD Monitoring [tos & sos]
Chesapeake Bay Program
Univ Maryland Chesapeake Biological Lab
Smithsonian Environmental Research Center
Thomas Point [TPLM2 C-MAN] – NOAA NBDC
River discharge - USGS Water Services
Bathymetry - NOAA NCEI
GIS & Satellite imagery
Natural Earth
Chesapeake Bay Program
NASA EOSDIS Worldview
NASA Visible Earth